Skip to content

fix moe gate dtype, fix tbo, fix fake dispatch #7825

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
zhyncs merged 5 commits into from
Jul 20, 2025
+3 −3
Merged

fix moe gate dtype, fix tbo, fix fake dispatch #7825

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
730b9cd
fix moe gate dtype, fix tbo, fix fake dispatch
Atream Jul 7, 2025
d501649
Merge branch 'main' into fix-MoEGate
Atream Jul 9, 2025
8c4010b
Merge branch 'main' into fix-MoEGate
zhyncs Jul 10, 2025
8f359f7
Merge branch 'main' into fix-MoEGate
Atream Jul 14, 2025
1e2e69a
Merge branch 'main' into fix-MoEGate
zhyncs Jul 15, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion python/sglang/srt/eplb/expert_location_dispatch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,7 +66,7 @@ def transform_select_experts_inputs(
info: Optional[ExpertLocationDispatchInfo],
):
if (info is not None) and (info.ep_dispatch_algorithm == "fake"):
router_logits = torch.randn_like(router_logits)
router_logits.uniform_(5, 10)
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Jul 7, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

Using magic numbers 5 and 10 directly in the code harms readability and maintainability. It would be better to define these as named constants with descriptive names (e.g., FAKE_DISPATCH_UNIFORM_MIN, FAKE_DISPATCH_UNIFORM_MAX) at the module level. This makes the purpose of these values clear and simplifies future modifications.

if correction_bias is not None:
correction_bias = torch.zeros_like(correction_bias)
return router_logits, correction_bias
Expand Down
2 changes: 1 addition & 1 deletion python/sglang/srt/layers/moe/topk.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -320,7 +320,7 @@ def biased_grouped_topk_gpu(
and is_power_of_two(correction_bias.shape[0])
):
topk_weights, topk_ids = moe_fused_gate(
gating_output,
gating_output.to(dtype=torch.float32),
correction_bias,
num_expert_group,
topk_group,
Expand Down
2 changes: 1 addition & 1 deletion python/sglang/srt/models/deepseek_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -219,7 +219,7 @@ def __init__(
)
if config.topk_method == "noaux_tc":
self.e_score_correction_bias = nn.Parameter(
torch.empty((config.n_routed_experts))
torch.empty((config.n_routed_experts), dtype=torch.float32)
)
else:
self.e_score_correction_bias = None
Expand Down
1 change: 1 addition & 0 deletions python/sglang/srt/two_batch_overlap.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -490,6 +490,7 @@ def filter_batch(
output_dict["spec_info"] = output_spec_info
for key in [
"forward_mode",
"is_extend_in_batch",
"return_logprob",
"req_to_token_pool",
"token_to_kv_pool",
Expand Down